00:00:00: And I would also relate this to my old friend Plato, namely the question is symbolic reasoning.

00:00:08: The foundation of intelligence?

00:00:10: That Plato said that it's a word of ideas as sort of the fundament and what we see in what you perceive as reflection on the world of ideas.

00:00:18: This is reversed!

00:00:20: It really is the other

00:00:21: way around And today is a very special day, and I mean you all will not know exactly which date it is because we record these episodes in advance.

00:00:37: But this was the birthday episode for Torsten so thank-you very much too that you still take time from our audience at For Me & For This Project on this very special Day.

00:00:49: Great!

00:00:49: We wanted to talk little bit about AI Because i mean...we have this sub series ongoing with few episodes more focused on this topic And I wanted to inquire a little bit about the relation between Go-fi, so good old fashioned AI and this more statistically based.

00:01:08: If you want to phrase it like that is large language models where we have these problems of hallucinations and stuff like that but off course on the other side which seems to be much more potent technology in some sense.

00:01:23: So sometimes say something Go-fi approach seems failed.

00:01:30: Why do you think so?

00:01:32: Okay, I'm a child of Go-Fi When i was an undergraduate student at TU Berlin in the nineteen eighties... ...I was really excited about AI and went to the German Spring School on AI, KI Frühe Schule And to the german workshop on AI even when there were students And I learned lots of things.

00:01:59: I mean, I learn lists with Pollock and...and i was really excited about this!

00:02:07: And maybe that wasn't the only one.

00:02:09: there's a big promise sort of bubble big bubble.

00:02:14: There were lots of investment in fifth generation computing stuff like this yeah?

00:02:22: It all would say it came.

00:02:24: It came to more or less nothing.

00:02:26: Also, expert systems were created and I was working for companies who worked in this area And more and more...I started have a bad feeling because i realized that the promises which are made they're not met with actual tools or programs, systems which would fulfill these promises.

00:02:54: So it was really a gap and I felt guilty about being involved in this overblown promise.

00:03:08: that is also the reason why I personally moved away from AI programming language with lambda calculus and so on.

00:03:20: Because this was a bit solid, yeah?

00:03:23: Whereas the AI stuff was like clearly above it And that's an idea of the old AI Was to base artificial intelligence on symbolic structures On symbolic logic epistemic logics, which take into account how moral people actually reason.

00:03:53: Logics of defaults and modality and so on.

00:03:57: that was all put in.

00:04:00: there were grammars which were much more evolved than the sort of Stannachomsky style grammars but had much more complicated structures to model actual human languages.

00:04:15: And tools were built based on these grammar-based language models, and it was quite depressing to see how they're performed because as long you stuck with the script or knew what this system could swallow normal sort of interaction, but if you put somebody who didn't know how to talk.

00:04:49: To the system in front it would completely fail was very brittle there.

00:04:55: so that was a situation and as I said also personally for me i mean It had some consequence even though they have jobs with AI.

00:05:08: In the end, for my academic career I decided that I didn't want to pursue this but instead do stuff which i'm still doing type theory, lambda calculus logic and so on.

00:05:24: Yeah?

00:05:25: First let me just say a lot of people would also consider this go-fi right.

00:05:29: it's AI used to be very broad area maybe capturing all of logic if you wanted to speak like but of course it's less applied.

00:05:39: maybe, is that?

00:05:40: No.

00:05:42: It's based on symbolic structures I mean the logics which were used and we had to discuss this a bit already.

00:05:58: they're philosophical logic epistemic logics, so modal logic which lots of modalities and... And so on.

00:06:08: But the idea was that we could model human reasoning by just making up an epistomic logic Which does all this default reasoning?

00:06:19: and it didn't work.

00:06:20: It really did not work.

00:06:21: I mean people are still trying but it doesn´t work Yeah?

00:06:30: to get through this philosophical underpinning, the insights we can gain from this failure.

00:06:39: But first of all let's go a bit more back into history.

00:06:44: and at the same time there were already people investigating neural networks to serve like, okay we have to model neurons and in this way you can model intelligence.

00:07:04: So it wasn't taking a serious... And that didn't produce any interesting results or maybe something?

00:07:11: Maybe I'm not fair but as far as i know It was an alternative to symbolic AI But it didn't produced As far as I know Any exciting result.

00:07:28: I guess the biggest lighthouse could be the recognition of written numbers, right?

00:07:34: The stuff where Lecombe was involved.

00:07:38: Right!

00:07:38: That's certainly something which fits into this... ...which is a starting point for his development.

00:07:49: and what then happened that made sort-of statistic based AI approach work, were some innovations or one hand processor speed and particularly parallel processing which was actually originally needed to graphic three-D simulations that scaled considerably.

00:08:19: And the other aspect is data sources via internet.

00:08:27: enormous growth of the internet, data sources were there in large supply.

00:08:33: And so I think that combination of these data sources and speed up computing created this revolution where we see now such large language models which I mean, the measure for AI was always a Turing test.

00:08:55: So Turing suggested how can we recognize artificial intelligence?

00:09:01: And he suggested that somebody would sit on a teletype writer and communicate with either someone else or an AI say whether this other partner was a human or an alien, if they kind of distinguish the AI from a human then that would be artificial intelligence.

00:09:27: And that's a test any LLM nowadays will easily pass right?

00:09:35: I mean If we stick little longer to these particular tests.

00:09:39: allow me two remarks.

00:09:40: one as a philosopher need to say off course original Turing Test is slightly different Because this test is unfair, right?

00:09:48: I as a human just need to be a human and the computer needs to look like a human.

00:09:53: So there's an asymmetry in Turing's tests was little more similar but of course still unfair for the computer.

00:10:01: um i think it is original.

00:10:03: uh man male person needed to convince somebody that they are female And the computer needed to convince.

00:10:13: So there's a little bit more distance between not just representing yourself, but that may be maybe the historic remark.

00:10:23: And yeah other thing is... A lot of approaches for the touring tests used what we might call like little tricks right?

00:10:33: The first expert systems got very well scores played weird roles being pretending to be

00:10:43: Caranult

00:10:44: paranoid or super focused on particular themes, US politics of football.

00:10:51: Or whatever are hard coded errors in typing and going deleting what was written at writing it back and things like that maybe?

00:11:02: That's kind of paradigmatic for the problems.

00:11:07: off this go fire approach because It so hard coded you really optimize local manner, and whenever you go too far outside of these locality your theory is simply not prepared.

00:11:22: And now even further away from this there's the quote of Gauss who had the approximation of prime numbers... ...and it was a very simple formula.

00:11:32: then there was specialized one with an error correction term things like that which were better for data they have so far.

00:11:40: But Gauss said, I don't care in the long run like this simple solution without extra tweaking will be correct one.

00:11:49: And that's maybe somehow what happened with modern AI systems?

00:11:54: Yes so if you look back then there was a competition between symbolic based AI and statistic-based AI And in a way, symbolic.

00:12:10: basically I really wanted to understand reasoning.

00:12:13: Where you could say the statistic based method at least from my point of view look like always lots of tinkering and trying things out.

00:12:27: sometimes there were cases where they think were much better than expected.

00:12:33: They had their own dynamics, you know like particularly the soft contextualization of an important aspect of LLMs created a much more so apparent intelligent behavior.

00:12:46: then it was expected right?

00:12:48: So um but the question is do we because we can now build uh AIs?

00:12:57: I mean systems which we can communicate with.

00:13:03: And they are very flexible and very, very fluent in the language use?

00:13:10: this is really quite impressive that I can write essays so there kind of have discussions or any topic and then also write programs right?

00:13:24: Display something which you can only call intelligence.

00:13:29: I mean, it's very difficult to deny that I am display intelligent behavior.

00:13:39: but the question is do we now understand intelligence?

00:13:43: And in a way... We don't right!

00:13:55: Do we really understand why, how this intelligence emerges or what is the substance of intelligence?

00:14:05: I would say you don't.

00:14:08: We know there's all these tinkering involved as an engineering approach where we manage to build something but we do not really understand it.

00:14:17: However i will tell that gives us a very important insight namely about the nature of intelligence.

00:14:27: And I would also relate this to my old friend Plato, namely as a question is symbolic reasoning?

00:14:41: The foundation?

00:14:43: And in a way I want to relate this, and i'm sure that I am getting hit now with the philosophy sticked by misquoting or misinterpreting.

00:14:53: But I would say that Plato said that the word of ideas is sort of the fundament... ...and what we perceive as reflection on the world of ideas.

00:15:06: This is reversed!

00:15:09: It's really the other way around the statistics, this fuzzy understanding of the world.

00:15:22: So I would say that it's a beversal.

00:15:29: The word, the immersion-word creates the words and ideas.

00:15:33: The idea is reflection on real worlds which we can approach via sort our senses, in a way our brains produce maybe similar due to evolution statistical analysis of processes and only then when this is completed or then we reflect upon our insights or understanding.

00:16:03: And make it precise!

00:16:04: That's where formal logic It's not fundamental, it is a starting point but reflection.

00:16:16: And I relate this to the failure of the symbolic AI namely that was guided by this idea.

00:16:29: we can reduce everything our thinking and reasoning through these symbolic processes.

00:16:40: processes are only the second step.

00:16:42: The first step is a sort of naive, emotion-based not reasoning or way to react to inputs and learn how to touch the hot plate again because you burn yourself all these evolutionary acquired abilities the word of ideas being created as a second step.

00:17:15: I guess i would say there's like, uh...a slight mix up there right?

00:17:19: Because that is the epistemic side.

00:17:21: how do we get to know and they're so-to speak maybe the ontologic site which has priority Which was their first?

00:17:28: And I mean for the ontologic side no idea I don't care So much!

00:17:33: I'm not doing metaphysics.

00:17:35: But

00:17:35: um but For the epistemics I think you are quite right.

00:17:40: And these emergent behaviors, something super interesting.

00:17:45: if we go a step back and look at LLMs just as language prediction tools then for me it was surprise because of course i know that world knowledge is so useful makes sense but its surprising

00:18:07: I mean, it has to do with ontology.

00:18:12: It's a question how does ontology emerges?

00:18:17: Is this the question whether ontology is fundamental... ...to understanding the world or whether in a way its secondary right?

00:18:29: And for LLMs its secondary.

00:18:32: but LLMs can create ontologies yeah But they're not built-in They emerge.

00:18:39: Scientology is emergent.

00:18:41: It's not the fundament, right?

00:18:44: Yeah and I mean it's impressive how much emergent behavior comes out of token prediction.

00:18:52: so i mean yeah.

00:18:55: but then it may be also interesting how much immersion behavior comes from this survival machine.

00:19:07: evolution of humans, I mean if you go like a little bit back and have some more simple organisms.

00:19:15: And look how they react in how they learn yeah?

00:19:18: Then you see this path from the simple organisms up to the brain is just evolutionary paths which creates this surprising thing!

00:19:33: The human brain was not designed.

00:19:35: designed, I mean to make mathematics or to write poems.

00:19:45: It was designed to survive and all this world of ideas.

00:19:53: really a side effect.

00:19:56: so maybe similar.

00:19:58: the LLMs were not originally designed.

00:20:04: Do what they can do now, in particular write programs.

00:20:06: Write proofs and have intelligent conversations with us.

00:20:10: They were much more had initially a much more narrow scope And as you say on the character recognition or whatever playing a game.

00:20:23: That was The first steps.

00:20:25: but But evolution of not the cultural evolution Of these RM's reflects The biological evolution of the human brain.

00:20:35: And it's similar that there is a surprising emergent structure which seems to come from nowhere once we build these complex systems based on statistical methods using learning, using lots of data sources We get in some kind of intelligent behavior I mean in a way.

00:21:04: And that's, I think it is very important for us to draw this conclusion when understanding reasoning and understand intelligent we really need start with the so statistic machine learning analysis of inputs So secondary, they're emergent.

00:21:35: I mean let me say two things.

00:21:36: the first is our brain Is not perfect for truth or optimization in many senses.

00:21:44: i mean sometimes whatever

00:21:45: to this.

00:21:46: okay

00:21:47: yeah maybe even less big terms.

00:21:51: if If I eat a whole bar of chocolate that is very good For my neanderthal brain it says oh good carbs fats.

00:21:59: you need that because it was a good behavior in the past, and prehistoric times when food was rare.

00:22:09: I had a poppy seed cake half an hour ago!

00:22:15: Yeah but there's this misalignment right?

00:22:18: In my brain between environment nowadays and the heuristics that used to be optimal for reproduction of whatever genes or alleles or what the whole population at large and doesn't matter too much.

00:22:35: And similar, I mean understanding that just because your brain says something it's not always optimal?

00:22:43: It also helps with these large language models.

00:22:47: you need to know.

00:22:48: understand when like being psychophantic is a good strategy if we want reward points from Sure.

00:22:58: We have another bias, we have confirmation bias.

00:23:01: so you know oh you need to be careful when these things produce language that says your idea is great and...

00:23:09: Oh I quite like

00:23:13: this!

00:23:15: My LLM told me it's a great point.

00:23:22: Then the other is, and maybe let me try to bridge to go this way.

00:23:26: But quickly

00:23:27: there's a comment about what you have just said This mismatch which has something to do with this mismatch between biological and cultural evolution right?

00:23:40: That cultural evolution goes much faster.

00:23:43: so we are ourselves changing our environment So that we're creating an environment for which we are not biologically tuned or even the design of our brain is not tuned to react in an appropriate way?

00:24:00: Yeah, I mean probably there...I'm very unsure about these things because i am very unsure how evolution exactly works.

00:24:09: There's this idea that biological things work faster than what you always thought because of epigenetics and

00:24:17: things like that.

00:24:18: But

00:24:19: as a rough idea, the environment moves very quickly.

00:24:23: Fifty years ago it looked different.

00:24:27: from now on...

00:24:28: There's no time for biology to adapt?

00:24:33: Yeah but let me go through this connection with GoFly.

00:24:37: I mean you could challenge this being a primer in two ways.

00:24:46: One would be something like the mathematics of machine learning, so maybe we can step back and may have great theory on all these things?

00:24:57: The other one is just having shortcomings of hallucinations or unexpected output which simply comes by being flexible and adversarial attacks.

00:25:10: stuff is now a possibility.

00:25:13: We need the other approach, at least as a guardrail and ideally yeah sure please

00:25:22: Yeah no I wanted to come actually sorry that's the next thing i want you two say.

00:25:26: And then i would call it a paradoxical synthesis side.

00:25:30: So we have this contradiction.

00:25:32: or there was this fight between competition Between symbolic AI and statistic based AI and clearly symbolic based, since statistic-based AI won the race.

00:25:46: However there's now an interesting opportunity to combine.

00:25:53: the two are actually essential to combine them too And we see this already here in our talk with Tudor that And if we ask an AI to produce a proof, then how can be sure that this proof is actually valid?

00:26:13: I mean it does in English and who would know.

00:26:17: So they adopted lean, exactly symbolic logic to validate the output of their AI.

00:26:28: It's also a thing I mentioned already.

00:26:36: We had this promise of formal verification that we could formally verify mathematical theorems, we can firmly verify systems but always said in principle because all the theorems were verified.

00:26:53: they're very much basic beginner level theorem And the systems were small and easy designed to be verifiable.

00:27:04: But we couldn't do this for contemporary mathematics, or real systems.

00:27:10: because yes in principle it would possible but in practice effort is just too big!

00:27:19: Now comes a statistic-based AI that actually makes it possible.

00:27:27: So in a way, statistic-based AI is now the enabling technology for formal logic.

00:27:34: Because I mean as we have seen the verification of Fermat-Lach's theorem on this Navier Stokes equation which was not completed or it wasn't accepted but that we will be able to do contemporary mathematics with AI.

00:28:04: And I think it would also enable us to do verification of real software systems as they are built, so this is a paradoxical synthesis because Statistic-based AI.

00:28:24: Statistic based AI wins, but in a way now enables the formal logic so symbolic logic together.

00:28:37: this is really interesting because we can have very efficient but also trustworthy AI and that's completely I would say The triumph of Fassi, statistically I enables the triumph of exact symbolic reasoning.

00:28:59: I mean i really like this conclusion and it also raises a lot questions for another premises.

00:29:05: because And

00:29:13: in life, right?

00:29:15: Yeah.

00:29:16: But the matter is more important than live.

00:29:18: I mean

00:29:20: come

00:29:22: on!

00:29:23: So we need to talk about that maybe in the next premises but let me just add... The other way around.

00:29:31: it's still also true because a lot of advanced capabilities if you work with your chatbot are now that this chatbot writes in Python Because language and programming is like an ability, again for me surprisingly well emerged from these models.

00:29:50: and then they are writing their own scripts.

00:29:53: And there's this classical example right how many Rs are in strawberry?

00:29:57: No idea!

00:29:58: How many words on this text?

00:29:59: no

00:29:59: idea?!

00:30:00: And if you ask it now write me a four hundred word text... You see the thinking of writing a python script counting the words.. Now that gets to four hundred words right which wasn't able Let me say two years ago, these things are so quickly developing.

00:30:17: No idea what exactly the premise is?

00:30:20: So What do you think?

00:30:22: shall we call it today for this premises and have another one focused on the changes of mathematical practice?

00:30:29: Yes I would maybe be go a bit more general.

00:30:35: Yeah,

00:30:37: that's also a big question.

00:30:39: Okay but for today we got to.

00:30:42: Hegel would be proud of us.

00:30:45: We go through the synthesis off classical go-fi and modern AI systems.

00:30:51: So thank you all for tuning in for buying us a coffee For becoming a channel member or commenting sharing liking subscribing All other things That helped us too grow so quickly.